Application of the genetic algorithm for agent-based modeling of the object behavior while leaving the flaming room

Abstract

The article deals with the implementation of the genetic algorithm for training and optimization of the neural network and its application to the tasks related to agent-based modelling of human behaviour. After the analysis of existing agent-based modelling programs, several drawbacks were noticed. The main problem of other systems for crowd modelling was the missing of information about the psychoemotional state of people, who are in crowd. According to the other sources, moods in crowd influence its behaviour the most; therefore, we decided to propose another method of creating more realistic crowd behaviour. The system that implements training of agents by selecting the most effective strategies of behaviour from the existing set of strategies using the genetic algorithm was proposed. In addition, this article highlights the detailed development of one agent behaviour module based on the neural network, which help the agent to navigate in the environment on condition of being trained enough. Due to created training method, it was mentioned, that training environment affects whole training process, and so several surveys were made at different environment configurations. As a base for the system, single-layer perceptron was chosen because of its simplicity and small number of connections. Final variant of agent casts seven rays into different vectors from his centre. If any other object interferes those rays, the distance to him comes as input to the neural network. Another two inputs take amount of distance to the exit and angle between agent and exit. Right rotate and left rotate forces of agent are used as outputs for neural network. Genetic algorithm was used for selecting the proper weights between the neurons. The most important indicator that some genome should be chosen for future generations is the fitness of that genome. The main purpose for the agent is finding the exit and in case of hitting the flame, it will fail. That is why agent will be rewarded for reaching the exit and get penalty for hitting the flame or staying in the room for too long. During the first launch, 30 initial genomes get involved with the values of weights generated at random. Each of them passes five iterations of training, after which the total gene is written to the database. By the results of current generation training, the five most adapted genomes that will create next generation will be selected. The following genomes remain in the population and form their modified copies by mitosis. Both genomes are divided on a specific point and exchange parts of the genetic code. That process forms two other genomes, which inherit parts of the parents. As for the training environment, there are four different exits from the room but only one of them opens randomly during the iteration. The implemented software system is used to research and train the agent model bypass the obstacles and get to the endpoint. The work of the neural system with a different structure was considered: one runs with the characteristics given above, in the other system configuration, the exit was only one and stayed in constant place, the agent had only four rays, and for staying in the room for more than five seconds he was rewarded. The development of the environment for such training is described and the results of training are presented for various environmental configurations. The main goal and mission of such approach implementation is using trained agents to develop a system for crowd behaviour modelling in the building, which was set on fire. Ref. 11, fig. 6, tabl. 4.